Explainer
Held-out test set
Examples kept aside for the final check, never trained on or tuned against.
Split your examples three ways. The training set is what the model learns from. The validation set is what you look at while choosing settings or checkpoints. The held-out test set is kept aside for the final check. It only works if nothing you decide is based on it: scikit-learn calls testing on training data "a methodological mistake", and tuning settings until the test score looks good leaks information too (scikit-learn.org). Keep related examples, such as turns of one conversation, in the same split, and score models before and after on exactly the same items. Once you keep the checkpoint with the best validation score, that score is no longer an unbiased final measure. The test set, which took no part in the choice, is the final check.
Going deeper
Paraphrased or translated copies of test items slip past simple duplicate filters, so deduplicate across splits with fuzzy matching (Yang et al. 2023, arxiv.org). Freeze the test file and record its version with every result; if it has been consulted too often, retire it and build a fresh one.
In the long read: The data decides what it learns. Every related term is in Fine-tuning, every concept.
Read next: Overfitting